The classification of high-dimensional gene expression data presents formidable computational and accuracy challenges, primarily stemming from inherent dimensionality, noise, and feature redundancy. To address these limitations, this study introduces GeKAN, a novel framework that synergistically integrates advanced feature selection methodologies (Boruta and mRMR) with Kolmogorov-Arnold Network (KAN) variants to optimise nonlinear classification. Our approach demonstrably enhances both predictive precision and computational efficiency. The experiment confirms that our proposal achieves substantial improvements in classification accuracy (elevating performance by 8– \(10\%\) ) while reducing computational time by approximately threefold compared to conventional methods. These results underscore the framework’s efficacy in mitigating dimensionality-related constraints, establishing it as a robust solution for high-throughput genomic data analysis.